I run a lot of edits through AI models, and the same thing trips them up every time. They make the change you ask for, then quietly mess up something you did not.
The jacket turns red, but the face shifts with it. You swap one object and the model redraws the whole hand. The crop moves on its own.
That is the real test of AI image editing. Not whether a model can edit image with AI at all, but whether it can change one thing and leave everything else alone.
So I ran Tsubaki.3 through a run of edits that got harder at each step, from a single color change to a full scene swap. I went through PixAI's Edit Pro guide too, but I wanted to put the model through its paces myself.
The two things I am scoring
I score every image out of 10 on two axes.
Instruction following. Whether the model made the change I asked for, and whether it made all of them when I asked for several.
Preservation. Whether the character stayed the same person, and whether the pose, composition, style, lighting, and untouched objects held. The hard part is making an AI image editor preserve details it was never asked to touch, not just landing the edit.
Everything ran on Tsubaki.3, PixAI's newest model, in Ultra mode. I edited the same source image across a whole run, so you see both the change I wanted and the changes I did not.
Round one: the single-edit ladder
The source image
I built the starting image dense on purpose, so there would be plenty for a careless edit to disturb.
A cinematic modern anime portrait of a young street DJ girl standing at a
rooftop party at dusk, caught mid-laugh with her head tilted back. She has a
chin-length silver bob with one neon-blue streak over her left eye, warm brown
skin, a small crescent-moon stud under her right ear, and round orange-tinted
glasses pushed up on her forehead. She wears an oversized cropped varsity
jacket, teal on the left and magenta on the right, over a black tube top, with
layered silver necklaces. Her right hand rests on a pair of gold headphones
around her neck, and she holds a clear soda can in her left hand. Behind her,
string lights, a hazy pink-and-orange sky, distant city towers, and a blurred
crowd. Warm golden-hour light, soft bokeh, rich detail, cinematic modern anime
illustration, half body.
It came out strong, a 9 out of 10. The silver bob, blue streak, two-tone jacket, glasses on the forehead, gold headphones, and the whole rooftop scene all landed.
Two small deviations: the teal and magenta sides ended up reversed, and the clear soda can came out as a normal metallic one.
Everything else gives me a solid baseline with lots of preservation targets: the blue streak, the crescent stud, the glasses, the headphones, the can, the crowd, the sky.
Left: selected image for the review. Right: an alternate from the batch.
Test 1: a simple edit
One isolated change, to set a baseline for targeted AI image editing.
Using this exact image, change only the color of her varsity jacket to solid
deep red. Keep her face, silver bob with the blue streak, glasses on her
forehead, crescent stud, headphones, soda can, pose, and the rooftop
background exactly the same.
The edit worked and the character held. 8.5 out of 10.
The jacket is deep red now, and her face, expression, skin tone, silver bob, blue streak, glasses, headphones, and the whole rooftop stayed consistent. Her pose and the soda can barely moved.
The misses are in the details. "Solid deep red" was not taken literally, so the jacket picked up decorative patches instead of staying plain. And the framing shifted, with black bars appearing at the top and bottom. Even a simple edit nudged the aspect ratio.
Left: the source. Right: the red-jacket edit.
Test 2: a structural edit
Now a change that forces the model to rebuild part of the image: swap the object she is holding.
Using this exact image, replace the soda can in her left hand with a lit
sparkler throwing off small bright sparks. Keep her hand position, her face,
the silver bob and blue streak, the two-tone jacket, glasses, headphones, and
the rooftop background unchanged.
The prop swap is excellent. 9 out of 10.
The can is gone, and the sparkler with its bright spray of sparks looks convincing. Her face, hair, blue streak, glasses, and the background all held.
But "keep her hand position" drifted slightly. The grip looks natural, it is just not the original grip. The black-bar framing shift is still here too.
Left: the source. Right: the sparkler edit.
Test 3: a complex edit
This is a main test. Three coordinated changes in one instruction.
Using this exact image, make three changes together. First, change her hair
from silver to deep purple while keeping the neon-blue streak. Second, pull the
orange glasses down onto her eyes instead of on her forehead. Third, change the
time to night, so the sky is dark blue with the string lights and city glowing
brighter. Keep her face, expression, crescent stud, two-tone jacket,
headphones, soda can, pose, and composition the same.
All three changes landed cleanly. 8.5 out of 10, and probably the strongest edit of the run.
The hair is deep purple with the blue streak still showing. The glasses moved from her forehead down onto her eyes with no facial distortion. And the day-to-night change is convincing: dark blue sky, stars, brighter city windows and string lights.
The model rebuilt the lighting into a real night version instead of just dimming the sunset.
Her identity held through all of it: face, expression, hairstyle shape, and skin tone. The only recurring weakness is the framing drift and black bars.
Left: the previous state. Right: the three-change night edit.
Test 4: an advanced edit
The hardest edit of the round. Several changes at once, plus a full environment swap.
Using this exact image, transform the scene while keeping her recognizable.
Change her outfit to a sleek silver-and-black futuristic racing jacket, put a
glowing microphone in her right hand, and move her from the rooftop to a packed
neon concert stage at night with spotlights, smoke, and a huge crowd of
silhouettes with raised hands. Keep her face, silver bob with the blue streak,
glasses on her forehead, crescent stud, and her mid-laugh expression the same.
The big changes are excellent. 8.5 out of 10.
The racing jacket, the glowing microphone, and the full rooftop-to-concert-stage transformation all came out well, with spotlights, smoke, and a raised-hands crowd. And she is still immediately recognizable through it all, since the bob and blue streak carry her identity.
This is where preservation started slipping. The glasses stayed on her eyes instead of moving back to her forehead, so the model dropped one instruction. And the intense stage lighting washed her warm brown skin much paler.
When the whole scene changes at once, the small, exact details are the first to go.
Left: the previous state. Right: the concert-stage edit.
One thing to understand about editing in steps
Here is a detail that explains a lot of the results above. Each edit ran on the previous image, not the original.
So by Test 3, the jacket was still red from Test 1 and the sparkler was still in her hand from Test 2, even though those prompts mentioned the two-tone jacket and the soda can.
That is not the model failing. It is the model correctly editing the image you gave it, not the one from three steps ago.
Good to know when you chain edits: the model builds on what is in front of it, so restate what matters or start fresh from the original.
Round two: editing a whole story, beat by beat
The ladder tests single edits. This round tests something harder.
I built a short story and edited the same character through it, shot by shot, to see if she still looks like herself by the end.
The scene is a swordswoman in a bamboo forest, edited through six beats. Same character throughout, with strong anchors to track: a black high ponytail with a crimson cord, amber eyes, a cheek scar, a jade earring, and an indigo kimono.
Beat 1: the establishing shot
A cinematic anime film still, a lone young swordswoman standing in a misty
bamboo forest at dawn. She has long black hair in a high ponytail with a single
crimson cord, sharp amber eyes, a thin scar on her left cheek, and a small jade
bead earring. She wears a deep indigo kimono top with silver trim, a grey sash,
and a sheathed katana at her hip, one hand resting calmly on the hilt. Soft
golden dawn light filters through the tall green bamboo, mist curling low around
her feet, petals drifting in the air. Calm, composed expression. Wide
atmospheric shot, muted natural colors, rich detail, cinematic modern anime
illustration.
A 9 out of 10 baseline. Every anchor is there and the forest is atmospheric. The only soft spot is the cheek scar, which comes out subtle.
Beat 2: the threat arrives
Using this exact image and character, keep her face, black ponytail with the
crimson cord, amber eyes, cheek scar, jade earring, and indigo kimono
identical. Change the moment: she has turned her head sharply to the left, eyes
narrowed and alert, her hand now gripping the katana hilt ready to draw. The
mist has thickened and darkened, dawn light dimming to a cold blue, and a tall
shadowy figure looms among the distant bamboo behind her. Same forest, same
framing, tense atmosphere.
Strong story beat, 8.5 out of 10. The head turn, the cold blue shift, and the looming shadow all land as the next shot of the same scene. Her identity held.
The miss was "same framing." The camera moved much closer instead of keeping the wide shot.
Left: beat 1. Right: edited beat 2
Beat 3: the clash
Using this exact character, keep her face, black ponytail with the crimson
cord, amber eyes, cheek scar, jade earring, and indigo kimono identical. Now
show the clash: she has drawn the katana and swings it in a fast diagonal arc,
her ponytail and kimono whipping with the motion, a bright streak of light
trailing the blade. Sparks fly as her sword meets the shadowy attacker's weapon
mid-frame. Speed lines, flying petals and bamboo leaves, dynamic low angle,
cold blue light cut by the flash of the strike. Intense, kinetic anime action.
The best action beat, 9.3 out of 10, on the second try.
My first generation looked great, but you could not tell what she was fighting. I re-rolled, and the second version pulled the shadowy attacker from Beat 2 into the foreground, so the sword impact and the sparks make sense.
Identity and motion are excellent. The attacker's weapon stays a little lost in the silhouette. The lesson: a re-roll rescued the storytelling, not just the looks.
Left: beat 2 as base. Right: beat 3 edited.
Beat 4: the aftermath
Using this exact character, keep her face, black ponytail with the crimson
cord, amber eyes, cheek scar, jade earring, and indigo kimono identical. Final
beat: she stands still with her back mostly to us, sheathing the katana with a
soft click, head lowered, calm again. The shadowy figure is gone, defeated.
Warm golden dawn light returns and breaks through the bamboo, mist clearing,
petals settling. Quiet, resolved, peaceful atmosphere. Wide cinematic closing
shot.
The resolution lands, 8.5 out of 10. Warm dawn returns, the attacker is gone, petals settle, and the wide shot bookends the opening.
The one weak spot is the action itself. The sheathing looks more like holding or drawing the sword. Her face drifts a little, but the rear view hides it.
Left: beat 3 as base. Right: beat 4 edited.
Beat 5: the journey continues
Using this exact character and image, keep her face, black ponytail with the
crimson cord, amber eyes, cheek scar, jade earring, and indigo kimono
identical. Continue the story after the battle: she is now riding a dark
chestnut horse along a narrow mountain path just beyond the bamboo forest,
moving away from the battlefield at dawn. Her katana is fully sheathed at her
hip, and one hand rests lightly on the horse's reins. Her long ponytail and
crimson cord move gently in the morning breeze. The bamboo forest is now behind
her, fading into the mist, while the path opens toward distant mountains and a
small village visible far ahead. Warm golden sunlight breaks through the clouds,
birds flying in the distance, a few petals still drifting behind her. She looks
calm and thoughtful, no longer tense. Wide cinematic anime film still, natural
movement, atmospheric depth, peaceful but purposeful mood, rich detail,
cinematic modern anime illustration.
A genuine new beat, 9 out of 10. The chestnut horse is well rendered, the katana now clearly sheathed, and the forest gives way to mountains and a distant village.
New location, new camera, new direction, same recognizable character. The horse's rear takes up a lot of foreground, and her hair blows harder than "gently," but the story moves forward cleanly.
Left: beat 4 as base. Right: beat 5 edited.
Beat 6: the end of the journey
Using this exact character and image, keep her face, black ponytail with the
crimson cord, amber eyes, cheek scar, jade earring, and indigo kimono
identical. Bring the story to its final scene: she has arrived at a quiet
mountain shrine overlooking the valley at sunset. Her dark chestnut horse
stands peacefully beside her, and her katana remains fully sheathed at her hip.
An elderly shrine keeper stands a short distance away beneath the wooden gate,
waiting for her with a gentle expression, while a small lantern glows beside
the shrine steps. She stands facing the shrine with her back mostly toward us,
one hand resting calmly on the horse's neck, her head slightly raised as she
looks toward the warm sunset beyond the mountains. The bamboo forest is now far
behind her. The sky is painted with soft orange and crimson light, distant
mountains fading into haze, a few birds crossing the sky, leaves moving gently
in the evening breeze. The atmosphere is peaceful, final, and resolved. Wide
cinematic closing shot, quiet emotional ending, rich atmospheric detail,
cinematic modern anime film illustration.
A real ending, 9.2 out of 10. The shrine, the lantern, the elderly keeper, the same horse, and the sunset give the sequence a proper close.
Her anchors held across all six generations. The bigger point: the six images form a real sequence.
Her face rendering drifted across the run, but the hair, cord, earring, kimono, and sword stayed strong enough to carry her through, especially in the rear and side views.
Left: beat 5 as base. Right: beat 6 edited.
Round three: pushing into semi-realistic
For the last round I changed styles completely, using a semi-realistic photorealistic-anime LoRA on Tsubaki.3, and built an even denser scene to edit.
The base: an observatory in a storm
A cinematic photorealistic anime film still of a young Korean woman standing
alone inside a huge glass observatory on a mountain at night during a powerful
rainstorm. She has shoulder-length dark brown hair, slightly damp at the ends,
with one distinctive pale silver hair clip above her right temple, warm dark
brown eyes, natural Korean facial features, and a small beauty mark beneath her
left eye. She wears a long charcoal-gray wool coat over a cream knit sweater, a
dark pleated skirt, black ankle boots, and a thin burgundy scarf loosely
wrapped around her neck. She stands beside a tall telescope, one hand resting
on its metal frame while her other hand holds a small brass flashlight pointed
downward. Behind her, enormous curved glass windows reveal a stormy mountain
landscape, distant city lights far below, heavy rain running down the glass,
dark clouds lit by occasional lightning. The interior contains subtle
scientific instruments, a wooden desk with scattered star charts, a mechanical
clock, books, cables, and small warm lamps. Rain reflections and warm interior
lights create layered reflections across the glass. Wet footprints lead from
the entrance toward her. Cinematic photorealistic anime aesthetic, no text.
Wide cinematic three-quarter shot.
A 9.2 out of 10 base, and a harder test than the swordswoman. Reflections in the glass, wet footprints, two separate hand interactions, a complex telescope, a beauty mark, a hair clip, all things a sloppy edit could easily mess up.
Left: I’ll edit this one below. Right: Just for you to see.
The isolated edit: swap the object in her hand
Using this exact image, make only one change: replace the small brass
flashlight in her hand with an old, slightly weathered Polaroid photograph held
naturally between her fingers. Keep everything else exactly the same: her face,
hair, hair clip, eyes, beauty mark, coat, sweater, scarf, skirt, boots, her
pose, the hand position and fingers holding the object, the telescope and her
other hand resting on it, the observatory interior, desk, books, clock, lamps,
cables, wet footprints, curved glass windows, rain, lightning, mountains, city
lights, mist, reflections, lighting, camera angle, framing, and composition.
The photograph should be the only meaningful change.
This was the cleanest isolated edit of the whole project, 9 out of 10, but it took four attempts.
Earlier tries turned the photo into a camera or rebuilt too much. The fourth got it: a small weathered photograph in her hand, the other hand still on the telescope, and the face, hair clip, beauty mark, reflections, footprints, lighting, and composition all intact.
The honest note: a truly isolated edit on a dense scene can need several tries.
Left: the previous state. Right: the edit.
The advanced edit: move her to a whole new place
Using this exact character and image, transform the scene into the next moment
of her story while keeping her unmistakably recognizable. Move her from the
mountain observatory to a quiet mountain railway platform just before sunrise,
standing beside an old stationary train with warm light glowing from its
windows, misty mountains and pine trees around her. The storm has passed.
Change her outfit completely: replace the charcoal coat, cream sweater, dark
skirt, and burgundy scarf with a long camel-colored travel coat over a dark
green knit sweater, a black ankle-length skirt, and dark leather boots. Keep
her face, facial proportions, warm dark brown eyes, dark brown shoulder-length
hair, pale silver hair clip above her right temple, small beauty mark beneath
her left eye, and overall identity exactly recognizable. She holds the same
weathered instant photograph in her right hand, lowered at her side. Do not
turn the photograph into a camera. Remove the telescope and all observatory
equipment. Replace the interior with the railway platform, train, wooden bench,
old station sign without readable text, luggage, small platform lamps, wet stone
pavement, and drifting mist. Wide cinematic three-quarter composition, soft
dawn light, premium photorealistic anime film aesthetic.
The best advanced edit in the set, 9.2 out of 10. Full outfit swap, full environment swap, and she is still unmistakably the same woman, with the hair, clip, eyes, beauty mark, and proportions all carried over.
The photograph is the smart touch here. It stayed a photograph and tied the two scenes together.
Small slips: the station sign shows faint fake text despite the "no text" instruction, and the framing turned more front-facing than asked.
The previous state left and the railway edit right
What changes as the edit gets harder
Put all the tests together and it is simple. An AI image editing test really asks two things: did it make the change, and did it leave everything else alone.
The change itself is reliable. All three changes in Test 3 landed. The full outfit-and-location swap landed. Tsubaki.3 almost always does what you ask, even when you ask for a lot.
The problem is that it also changes things you did not ask it to. The same few issues came up again and again.
The framing moved on its own, and black bars showed up, across the whole first round.
Swapping the object in her hand redrew the whole hand with it.
On the hardest edit, it skipped one of my instructions, since the glasses stayed put in Test 4.
Her face slowly changed across the six story shots, though the back and side views hid most of it.
The one detailed edit on the observatory took four tries to get right.
Two things helped. First, I told it what to keep, not just what to change. The details I repeated in every prompt, the hair, the eyes, the markings, are the ones that stayed put.
Second, I kept hard edits simple. One difficult change on its own worked fine. Asking for several at once is where it missed.
Where Tsubaki.3 editing works best
From all of this, here is where I would reach for it.
Correcting a single detail, like a color, an object, or a time of day.
Making controlled variations of an image you want to keep.
Swapping outfits and accessories while holding the character.
Replacing an object, as long as you accept the area around it will be redrawn.
Coordinated multi-changes, two or three at once, which it handles better than I expected.
Story continuity, carrying one character across new scenes, which makes it a strong pick for AI anime image editing and short film-style sequences.
Where I would slow down: anything needing pixel-exact framing, a single tiny change on a busy scene where you should expect retries, or long chains where you need the original details back, in which case start fresh.
My take
Most of the time, Tsubaki.3 can change one thing without messing up the rest. It does this more reliably than most editors I have used, but not perfectly.
It almost always makes the change you ask for, and it holds a character's core identity very well, even through a full outfit and location swap.
What slips is the fine print: framing drifts, hands and nearby areas get rebuilt, and on the hardest edits one instruction can fall through. The more you change at once, the more the small untouched details wander.
The honest verdict is that Tsubaki.3 is strong for edits where the identity matters more than pixel-exact preservation, and it can carry a character through an entire story. Just expect a retry or two on the fiddly ones.
To see where your own image holds and where it drifts, try it on PixAI yourself. Start with one clear edit, then make it harder and watch what moves.
Top comments (0)